[InstCombine] Remove assumptions about int having 32 bits

[msebor]

Summary

The patch submitted in D129915 and subsequently committed in rG0dcfe7aa35cd exposes a number of assumptions in SimplifyLibcalls.cpp about int having 32 bits. Prior to the change the assumptions, although strictly unsafe, were benign even on 16-bit targets because the validation in TargetLibraryInfoImpl::isValidProtoForLibFunc enforced the corresponding restriction. The change has corrected the function to consider valid on 16-bit targets those functions in whose declarations the int return type or argument corresponds to i16, but it did not update the corresponding call handlers in SimplifyLibcalls.cpp. The end result is that some valid C code that compiled successfully before the change (albeit with suboptimal output) now causes an abort in the handler when targeting 16-bit environments (AVR or MSP430 in tree).

As a follow-up to D129915, this change removes these 32-bit assumptions to let the same code compile successfully even for 16-bit targets, and to also enable the same optimizing transformations for those targets as for 32-bit ones.

I've tested this patch on x86_64-linux.

[uabelho]

I've verified that this patch fixes the crash for ffs that we saw for a target with 16bit int. Thanks!
I have no idea about test coverage though, I think we're pretty bad at testing lib calls for our own compiler. The ffs crash was the only I saw.

So all in all, if the original patch caused other mismatches between the checks and later expectations I have no clue about really.

But this seems to be a step in the right direction.

[msebor]

Thanks for the confirmation! I've CC'd @efriedma who has reviewed changes in this in this area before. If there are no other comments I'll plan on installing the patch later this week to resolve the ffs abort.

[bjope]

I still think it is hard to understand for which functions isValidProtoForLibFunc was modified in the earlier commit that both refactored and modified the behavior in the same commit. For an out-of-tree target maintainer I need to understand how this impacted our downstream modifications to isValidProtoForLibFunc, while at the same time understanding if the changes impact our target (which isn't just having 16-bit ints, but for example char is also 16-bits, etc). Anyway, now I guess the only way is forward (even if other out-of-tree target maintainers might end up having the same trouble when rebasing later), so I guess I'll stop dwelling about that.

LGTM, as this seem to solve problems seen with D129915.

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

2712

I do not fully understand the part with "Putchar converts it to unsigned char regardless". Is that referring to the libc implementation , or llvm::emitPutChar, or what?

Not sure if we can deal with this better regardless of the answer. I guess this would work for in-tree targets and "normal" compiler hosts. But unsigned char on the host may have a different size compared to size of unsigned char in the runtime libc implementation for the target. Since the getConstantStringInfo call above assume that a char is 8 bits (at least the in-tree version of getConstantStringInfo) and the compiler probably is built for a host with 8 bit char, then at least the compiler code match up here.

(FWIW, our OOT target with 16-bit char is now guarded by getConstantStringInfo returning false, while in the past I think isValidProtoForLibFunc might have guarded from even calling LibCallSimplifier::optimizePuts and LibCallSimplifier::optimizePrintFString.)

[msebor]

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

2712

The putchar comment refers to the libc implementation of the function on the target. The host putchar shouldn't come into play here, but I've seen the assumption that char is 8 bits wide on both the host as on the target hardwired into transformations that don't depend on getConstantStringInfo, such as optimizeMemChr and optimizeStrChr (I'd be surprised if those were the only ones). If other sizes are meant to be supported these assumptions should be reviewed and if necessary removed. I don't have access to such targets but I'm willing to help with that.

llvm/lib/Transforms/Utils/BuildLibCalls.cpp

Value *llvm::emitBinaryFloatFnCall(Value *Op1, Value *Op2,
   Module *M = B.GetInsertBlock()->getModule();
   LibFunc TheLibFunc;
   StringRef Name = getFloatFn(M, TLI, Op1->getType(), DoubleFn, FloatFn,
                               LongDoubleFn, TheLibFunc);
 
   return emitBinaryFloatFnCallHelper(Op1, Op2, TheLibFunc, Name, B, Attrs, TLI);
 }
 
+// Emit a call to putchar(int) with Char as the argument.  Char must have
+// the same precision as int, which need not be 32 bits.
 Value *llvm::emitPutChar(Value *Char, IRBuilderBase &B,
                          const TargetLibraryInfo *TLI) {
   Module *M = B.GetInsertBlock()->getModule();
   if (!isLibFuncEmittable(M, TLI, LibFunc_putchar))
     return nullptr;
 
+  Type *Ty = Char->getType();
   StringRef PutCharName = TLI->getName(LibFunc_putchar);
-  FunctionCallee PutChar = getOrInsertLibFunc(M, *TLI, LibFunc_putchar,
-                                              B.getInt32Ty(), B.getInt32Ty());
+  FunctionCallee PutChar = getOrInsertLibFunc(M, *TLI, LibFunc_putchar, Ty, Ty);
   inferNonMandatoryLibFuncAttrs(M, PutCharName, *TLI);
-  CallInst *CI = B.CreateCall(PutChar,
-                              B.CreateIntCast(Char,
-                              B.getInt32Ty(),
-                              /*isSigned*/true,
-                              "chari"),
-                              PutCharName);
+  CallInst *CI = B.CreateCall(PutChar, Char, PutCharName);
 
   if (const Function *F =
           dyn_cast<Function>(PutChar.getCallee()->stripPointerCasts()))
     CI->setCallingConv(F->getCallingConv());
   return CI;
 }
 
 Value *llvm::emitPutS(Value *Str, IRBuilderBase &B,

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

void LibCallSimplifier::classifyArgUse(
   }
 }
 
 //===----------------------------------------------------------------------===//
 // Integer Library Call Optimizations
 //===----------------------------------------------------------------------===//
 
 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilderBase &B) {
-  // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
+  // All variants of ffs return int which need not be 32 bits wide.
+  // ffs{,l,ll}(x) -> x != 0 ? (int)llvm.cttz(x)+1 : 0
+  Type *RetType = CI->getType();
   Value *Op = CI->getArgOperand(0);
   Type *ArgType = Op->getType();
   Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
                                           Intrinsic::cttz, ArgType);
   Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
   V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
-  V = B.CreateIntCast(V, B.getInt32Ty(), false);
+  V = B.CreateIntCast(V, RetType, false);
 
   Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
-  return B.CreateSelect(Cond, V, B.getInt32(0));
+  return B.CreateSelect(Cond, V, ConstantInt::get(RetType, 0));
 }
 
 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilderBase &B) {
-  // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
+  // All variants of fls return int which need not be 32 bits wide.
+  // fls{,l,ll}(x) -> (int)(sizeInBits(x) - llvm.ctlz(x, false))
   Value *Op = CI->getArgOperand(0);
   Type *ArgType = Op->getType();
   Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
                                           Intrinsic::ctlz, ArgType);
   Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
   V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
                   V);
   return B.CreateIntCast(V, CI->getType(), false);
 }
 
 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilderBase &B) {
   // abs(x) -> x <s 0 ? -x : x
   // The negation has 'nsw' because abs of INT_MIN is undefined.
   Value *X = CI->getArgOperand(0);
   Value *IsNeg = B.CreateIsNeg(X);
   Value *NegX = B.CreateNSWNeg(X, "neg");
   return B.CreateSelect(IsNeg, NegX, X);
 }
 
 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilderBase &B) {
   // isdigit(c) -> (c-'0') <u 10
   Value *Op = CI->getArgOperand(0);
-  Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
-  Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
+  Type *ArgType = Op->getType();
+  Op = B.CreateSub(Op, ConstantInt::get(ArgType, '0'), "isdigittmp");
+  Op = B.CreateICmpULT(Op, ConstantInt::get(ArgType, 10), "isdigit");
   return B.CreateZExt(Op, CI->getType());
 }
 
 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilderBase &B) {
   // isascii(c) -> c <u 128
   Value *Op = CI->getArgOperand(0);
-  Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
+  Type *ArgType = Op->getType();
+  Op = B.CreateICmpULT(Op, ConstantInt::get(ArgType, 128), "isascii");
   return B.CreateZExt(Op, CI->getType());
 }
 
 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilderBase &B) {
   // toascii(c) -> c & 0x7f
   return B.CreateAnd(CI->getArgOperand(0),
                      ConstantInt::get(CI->getType(), 0x7F));
 }
if (FormatStr.empty()) // Tolerate printf's declared void.
     return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
 
   // Do not do any of the following transformations if the printf return value
   // is used, in general the printf return value is not compatible with either
   // putchar() or puts().
   if (!CI->use_empty())
     return nullptr;
 
+  Type *IntTy = CI->getType();
   // printf("x") -> putchar('x'), even for "%" and "%%".
-  if (FormatStr.size() == 1 || FormatStr == "%%")
-    return copyFlags(*CI, emitPutChar(B.getInt32(FormatStr[0]), B, TLI));
+  if (FormatStr.size() == 1 || FormatStr == "%%") {
+    // Convert the character to unsigned char before passing it to putchar
+    // to avoid host-specific sign extension in the IR.  Putchar converts
+    // it to unsigned char regardless.
+    Value *IntChar = ConstantInt::get(IntTy, (unsigned char)FormatStr[0]);
+    return copyFlags(*CI, emitPutChar(IntChar, B, TLI));
+  }
 
   // Try to remove call or emit putchar/puts.
   if (FormatStr == "%s" && CI->arg_size() > 1) {
     StringRef OperandStr;
     if (!getConstantStringInfo(CI->getOperand(1), OperandStr))
       return nullptr;
     // printf("%s", "") --> NOP
     if (OperandStr.empty())
       return (Value *)CI;
     // printf("%s", "a") --> putchar('a')
-    if (OperandStr.size() == 1)
-      return copyFlags(*CI, emitPutChar(B.getInt32(OperandStr[0]), B, TLI));
+    if (OperandStr.size() == 1) {
+      // Convert the character to unsigned char before passing it to putchar
+      // to avoid host-specific sign extension in the IR.  Putchar converts
+      // it to unsigned char regardless.
+      Value *IntChar = ConstantInt::get(IntTy, (unsigned char)OperandStr[0]);
+      return copyFlags(*CI, emitPutChar(IntChar, B, TLI));
+    }
     // printf("%s", str"\n") --> puts(str)
     if (OperandStr.back() == '\n') {
       OperandStr = OperandStr.drop_back();
       Value *GV = B.CreateGlobalString(OperandStr, "str");
       return copyFlags(*CI, emitPutS(GV, B, TLI));
     }
     return nullptr;
   }
 
   // printf("foo\n") --> puts("foo")
   if (FormatStr.back() == '\n' &&
       !FormatStr.contains('%')) { // No format characters.
     // Create a string literal with no \n on it.  We expect the constant merge
     // pass to be run after this pass, to merge duplicate strings.
     FormatStr = FormatStr.drop_back();
     Value *GV = B.CreateGlobalString(FormatStr, "str");
     return copyFlags(*CI, emitPutS(GV, B, TLI));
   }
 
   // Optimize specific format strings.
   // printf("%c", chr) --> putchar(chr)
   if (FormatStr == "%c" && CI->arg_size() > 1 &&
-      CI->getArgOperand(1)->getType()->isIntegerTy())
-    return copyFlags(*CI, emitPutChar(CI->getArgOperand(1), B, TLI));
+      CI->getArgOperand(1)->getType()->isIntegerTy()) {
+    // Convert the argument to the type expected by putchar, i.e., int, which
+    // need not be 32 bits wide but which is the same as printf's return type.
+    Value *IntChar = B.CreateIntCast(CI->getArgOperand(1), IntTy, false);
+    return copyFlags(*CI, emitPutChar(IntChar, B, TLI));
+  }
 
   // printf("%s\n", str) --> puts(str)
   if (FormatStr == "%s\n" && CI->arg_size() > 1 &&
       CI->getArgOperand(1)->getType()->isPointerTy())
     return copyFlags(*CI, emitPutS(CI->getArgOperand(1), B, TLI));
   return nullptr;
 }
 

 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilderBase &B) {
   annotateNonNullNoUndefBasedOnAccess(CI, 0);
   if (!CI->use_empty())
     return nullptr;
 
   // Check for a constant string.
   // puts("") -> putchar('\n')
   StringRef Str;
-  if (getConstantStringInfo(CI->getArgOperand(0), Str) && Str.empty())
-    return copyFlags(*CI, emitPutChar(B.getInt32('\n'), B, TLI));
+  if (getConstantStringInfo(CI->getArgOperand(0), Str) && Str.empty()) {
+    // putchar takes an argument of the same type as puts returns, i.e.,
+    // int, which need not be 32 bits wide.
+    Type *IntTy = CI->getType();
+    return copyFlags(*CI, emitPutChar(ConstantInt::get(IntTy, '\n'), B, TLI));
+  }
 
   return nullptr;
 }
 
 Value *LibCallSimplifier::optimizeBCopy(CallInst *CI, IRBuilderBase &B) {
   // bcopy(src, dst, n) -> llvm.memmove(dst, src, n)
   return copyFlags(*CI, B.CreateMemMove(CI->getArgOperand(1), Align(1),
                                         CI->getArgOperand(0), Align(1),

llvm/test/Transforms/InstCombine/ffs-i16.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+;
+; Test that the ffs library call simplifier works correctly even for
+; targets with 16-bit int.
+;
+; RUN: opt < %s -mtriple=avr-linux -passes=instcombine -S | FileCheck %s
+; RUN: opt < %s -mtriple=msp430-linux -passes=instcombine -S | FileCheck %s
+
+declare i16 @ffs(i16)
+
+declare void @sink(i16)
+
+
+define void @fold_ffs(i16 %x) {
+; CHECK-LABEL: @fold_ffs(
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    [[CTTZ:%.*]] = call i16 @llvm.cttz.i16(i16 [[X:%.*]], i1 true), !range [[RNG0:![0-9]+]]
+; CHECK-NEXT:    [[TMP1:%.*]] = add nuw nsw i16 [[CTTZ]], 1
+; CHECK-NEXT:    [[DOTNOT:%.*]] = icmp eq i16 [[X]], 0
+; CHECK-NEXT:    [[TMP2:%.*]] = select i1 [[DOTNOT]], i16 0, i16 [[TMP1]]
+; CHECK-NEXT:    call void @sink(i16 [[TMP2]])
+; CHECK-NEXT:    ret void
+;
+  %n0 = call i16 @ffs(i16 0)
+  call void @sink(i16 %n0)
+
+  %n1 = call i16 @ffs(i16 1)
+  call void @sink(i16 %n1)
+
+  %nx = call i16 @ffs(i16 %x)
+  call void @sink(i16 %nx)
+
+  ret void
+}

llvm/test/Transforms/InstCombine/fls-i16.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+;
+; Test that the fls library call simplifier works correctly even for
+; targets with 16-bit int.  Although fls is available on a number of
+; targets it's supported (hardcoded as available) only on FreeBSD.
+;
+; RUN: opt < %s -mtriple=avr-freebsd -passes=instcombine -S | FileCheck %s
+; RUN: opt < %s -mtriple=msp430-freebsd -passes=instcombine -S | FileCheck %s
+
+declare i16 @fls(i16)
+
+declare void @sink(i16)
+
+
+define void @fold_fls(i16 %x) {
+; CHECK-LABEL: @fold_fls(
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    [[CTLZ:%.*]] = call i16 @llvm.ctlz.i16(i16 [[X:%.*]], i1 false), !range [[RNG0:![0-9]+]]
+; CHECK-NEXT:    [[TMP1:%.*]] = sub nuw nsw i16 16, [[CTLZ]]
+; CHECK-NEXT:    call void @sink(i16 [[TMP1]])
+; CHECK-NEXT:    ret void
+;
+  %n0 = call i16 @fls(i16 0)
+  call void @sink(i16 %n0)
+
+  %n1 = call i16 @fls(i16 1)
+  call void @sink(i16 %n1)
+
+  %nx = call i16 @fls(i16 %x)
+  call void @sink(i16 %nx)
+
+  ret void
+}

llvm/test/Transforms/InstCombine/isascii-i16.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; Test that the isascii library call simplifier works correctly even for
+; targets with 16-bit int.
+;
+; RUN: opt < %s -mtriple=avr-freebsd -passes=instcombine -S | FileCheck %s
+; RUN: opt < %s -mtriple=msp430-linux -passes=instcombine -S | FileCheck %s
+
+declare i16 @isascii(i16)
+
+declare void @sink(i16)
+
+
+define void @fold_isascii(i16 %c) {
+; CHECK-LABEL: @fold_isascii(
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    [[ISASCII:%.*]] = icmp ult i16 [[C:%.*]], 128
+; CHECK-NEXT:    [[TMP1:%.*]] = zext i1 [[ISASCII]] to i16
+; CHECK-NEXT:    call void @sink(i16 [[TMP1]])
+; CHECK-NEXT:    ret void
+;
+  %i0 = call i16 @isascii(i16 0)
+  call void @sink(i16 %i0)
+
+  %i1 = call i16 @isascii(i16 1)
+  call void @sink(i16 %i1)
+
+  %i127 = call i16 @isascii(i16 127)
+  call void @sink(i16 %i127)
+
+  %i128 = call i16 @isascii(i16 128)
+  call void @sink(i16 %i128)
+
+  %i255 = call i16 @isascii(i16 255)
+  call void @sink(i16 %i255)
+
+  %i256 = call i16 @isascii(i16 256)
+  call void @sink(i16 %i256)
+
+  ; Fold isascii(INT_MAX) to 0.  The call is valid with all int values.
+  %imax = call i16 @isascii(i16 32767)
+  call void @sink(i16 %imax)
+
+  %uimax = call i16 @isascii(i16 65535)
+  call void @sink(i16 %uimax)
+
+  %ic = call i16 @isascii(i16 %c)
+  call void @sink(i16 %ic)
+
+  ret void
+}

llvm/test/Transforms/InstCombine/isdigit-i16.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; Test that the isdigit library call simplifier works correctly even for
+; targets with 16-bit int.
+;
+; RUN: opt < %s -mtriple=avr-linux -passes=instcombine -S | FileCheck %s
+; RUN: opt < %s -mtriple=msp430-freebsd -passes=instcombine -S | FileCheck %s
+
+declare i16 @isdigit(i16)
+
+declare void @sink(i16)
+
+define void @fold_isdigit(i16 %c) {
+; CHECK-LABEL: @fold_isdigit(
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    call void @sink(i16 1)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    call void @sink(i16 0)
+; CHECK-NEXT:    [[ISDIGITTMP:%.*]] = add i16 [[C:%.*]], -48
+; CHECK-NEXT:    [[ISDIGIT:%.*]] = icmp ult i16 [[ISDIGITTMP]], 10
+; CHECK-NEXT:    [[TMP1:%.*]] = zext i1 [[ISDIGIT]] to i16
+; CHECK-NEXT:    call void @sink(i16 [[TMP1]])
+; CHECK-NEXT:    ret void
+;
+  %i0 = call i16 @isdigit(i16 0)
+  call void @sink(i16 %i0)
+
+  %i1 = call i16 @isdigit(i16 1)
+  call void @sink(i16 %i1)
+
+  ; Fold isdigit('/') to 0.
+  %i47 = call i16 @isdigit(i16 47)
+  call void @sink(i16 %i47)
+
+; Fold isdigit('0') to 1.
+  %i48 = call i16 @isdigit(i16 48)
+  call void @sink(i16 %i48)
+
+  ; Fold isdigit('1') to 1.
+  %i49 = call i16 @isdigit(i16 49)
+  call void @sink(i16 %i49)
+
+  ; Fold isdigit('9') to 1.
+  %i57 = call i16 @isdigit(i16 57)
+  call void @sink(i16 %i57)
+
+  ; Fold isdigit(':') to 0.
+  %i58 = call i16 @isdigit(i16 58)
+  call void @sink(i16 %i58)
+
+  %i127 = call i16 @isdigit(i16 127)
+  call void @sink(i16 %i127)
+
+  %i128 = call i16 @isdigit(i16 128)
+  call void @sink(i16 %i128)
+
+  %i255 = call i16 @isdigit(i16 255)
+  call void @sink(i16 %i255)
+
+  ; Fold isdigit(256) to 0.  The argument is required to be representable
+  ; in unsigned char but it's a common mistake to call the function with
+  ; other arguments and it's arguably safer to fold such calls than to
+  ; let the library call return an arbitrary value or crash.
+  %i256 = call i16 @isdigit(i16 256)
+  call void @sink(i16 %i256)
+
+  ; Same as above.
+  %imax = call i16 @isdigit(i16 32767)
+  call void @sink(i16 %imax)
+
+  %ic = call i16 @isdigit(i16 %c)
+  call void @sink(i16 %ic)
+
+  ret void
+}

llvm/test/Transforms/InstCombine/printf-i16.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+;
+; RUN: opt < %s -mtriple=avr-freebsd -passes=instcombine -S | FileCheck %s
+; RUN: opt < %s -mtriple=msp430-linux -passes=instcombine -S | FileCheck %s
+;
+; Verify that the puts to putchar transformation works correctly even for
+; targets with 16-bit int.
+
+declare i16 @putchar(i16)
+declare i16 @puts(i8*)
+
+@s1 = constant [2 x i8] c"\01\00"
+@s7f = constant [2 x i8] c"\7f\00"
+@s80 = constant [2 x i8] c"\80\00"
+@sff = constant [2 x i8] c"\ff\00"
+
+@pcnt_c = constant [3 x i8] c"%c\00"
+@pcnt_s = constant [3 x i8] c"%s\00"
+
+declare i16 @printf(i8*, ...)
+
+; Verfify that the three printf to putchar transformations all result
+; in the same output for calls with equivalent arguments.
+
+define void @xform_printf(i8 %c8, i16 %c16) {
+; CHECK-LABEL: @xform_printf(
+; CHECK-NEXT:    [[PUTCHAR:%.*]] = call i16 @putchar(i16 1)
+; CHECK-NEXT:    [[PUTCHAR1:%.*]] = call i16 @putchar(i16 1)
+; CHECK-NEXT:    [[PUTCHAR2:%.*]] = call i16 @putchar(i16 1)
+; CHECK-NEXT:    [[PUTCHAR3:%.*]] = call i16 @putchar(i16 127)
+; CHECK-NEXT:    [[PUTCHAR4:%.*]] = call i16 @putchar(i16 127)
+; CHECK-NEXT:    [[PUTCHAR5:%.*]] = call i16 @putchar(i16 127)
+; CHECK-NEXT:    [[PUTCHAR6:%.*]] = call i16 @putchar(i16 128)
+; CHECK-NEXT:    [[PUTCHAR7:%.*]] = call i16 @putchar(i16 128)
+; CHECK-NEXT:    [[PUTCHAR8:%.*]] = call i16 @putchar(i16 128)
+; CHECK-NEXT:    [[PUTCHAR9:%.*]] = call i16 @putchar(i16 255)
+; CHECK-NEXT:    [[PUTCHAR10:%.*]] = call i16 @putchar(i16 255)
+; CHECK-NEXT:    [[PUTCHAR11:%.*]] = call i16 @putchar(i16 255)
+; CHECK-NEXT:    [[TMP1:%.*]] = zext i8 [[C8:%.*]] to i16
+; CHECK-NEXT:    [[PUTCHAR12:%.*]] = call i16 @putchar(i16 [[TMP1]])
+; CHECK-NEXT:    [[PUTCHAR13:%.*]] = call i16 @putchar(i16 [[C16:%.*]])
+; CHECK-NEXT:    ret void
+;
+  %ppcnt_c = getelementptr [3 x i8], [3 x i8]* @pcnt_c, i32 0, i32 0
+  %ppcnt_s = getelementptr [3 x i8], [3 x i8]* @pcnt_s, i32 0, i32 0
+
+  %ps1 = getelementptr [2 x i8], [2 x i8]* @s1, i32 0, i32 0
+  call i16 (i8*, ...) @printf(i8* %ps1)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_c, i16 1)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_s, i8* %ps1)
+
+  %ps7f = getelementptr [2 x i8], [2 x i8]* @s7f, i32 0, i32 0
+  call i16 (i8*, ...) @printf(i8* %ps7f)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_c, i16 127)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_s, i8* %ps7f)
+
+  %ps80 = getelementptr [2 x i8], [2 x i8]* @s80, i32 0, i32 0
+  call i16 (i8*, ...) @printf(i8* %ps80)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_c, i16 128)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_s, i8* %ps80)
+
+  %psff = getelementptr [2 x i8], [2 x i8]* @sff, i32 0, i32 0
+  call i16 (i8*, ...) @printf(i8* %psff)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_c, i16 255)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_s, i8* %psff)
+
+; The i8 argument to printf can be either zero-extended or sign-extended
+; when passed to putchar which then converts it to unsigned char.
+  call i16 (i8*, ...) @printf(i8* %ppcnt_c, i8 %c8)
+  call i16 (i8*, ...) @printf(i8* %ppcnt_c, i16 %c16)
+  ret void
+}

llvm/test/Transforms/InstCombine/puts-i16.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+;
+; RUN: opt < %s -mtriple=avr-linux -passes=instcombine -S | FileCheck %s
+; RUN: opt < %s -mtriple=msp430-freebsd -passes=instcombine -S | FileCheck %s
+;
+; Test that the puts to putchar transformation works correctly even for
+; targets with 16-bit int.
+
+declare i16 @putchar(i16)
+declare i16 @puts(i8*)
+
+@empty = constant [1 x i8] c"\00"
+
+define void @xform_puts(i16 %c) {
+; CHECK-LABEL: @xform_puts(
+; CHECK-NEXT:    [[PUTCHAR:%.*]] = call i16 @putchar(i16 10)
+; CHECK-NEXT:    ret void
+;
+; Transform puts("") to putchar("\n").
+  %s = getelementptr [1 x i8], [1 x i8]* @empty, i32 0, i32 0
+  call i16 @puts(i8* %s)
+
+  ret void
+}